Mamba-enhanced disease semantic knowledge graph for interpretable automatic ICD coding

Automatic ICD coding refers to the process of using artificial intelligence methods to automatically extract information related to diseases, symptoms, diagnoses, treatments, and other relevant details from electronic health records, and convert it into codes that comply with the International Classification of Diseases (ICD) standard. Automatic ICD coding technology has been gradually improved with the advancement of deep learning, but in practical deployment, it still faces challenges such as inconsistent semantics, ambiguous labels, and limited interpretability. To address these issues, we propose a novel automatic ICD coding framework MKHCNet (Mamba-Knowledge-HPLA-ContraNorm Network) which integrates unstructured clinical knowledge representation, long-range dependency modeling, and contrastive normalization techniques to enhance coding performance. Specifically, we construct a disease semantic knowledge graph to enrich ICD label representations, employ the Mamba network to capture cross-domain dependencies, apply the ContraNorm module to enhance label separability, and propose the Hierarchical Position Label Attention (HPLA) mechanism to achieve fine-grained, attention-based interpretability. Finally, with the purpose of capturing complex nonlinear relationships more effectively and better adapting to complex patterns in EHR data, FastKAN acts as a classifier and utilizes radial basis function (RBF) for feature transformation. We conducted systematic experiments on the benchmark datasets MIMIC-FULL and MIMIC-50. The experimental results show that MKHCNet improves MaAUC and P@8 by 2.1% and 0.3% on MIMIC-FULL respectively compared with the best existing mainstream model. Furthermore, case studies demonstrate that the model is able to effectively identify complex semantic cues and provide strong clinical interpretability.